This patchset attempts to fix https://bugs.launchpad.net/qemu/+bug/1823998
which reports that we don't handle kernels larger than 128MB
correctly, because we allow the initrd to be placed over the
tail end of the kernel. AArch64 kernel Image files (since v3.17)
report the total size they require (including any BSS area that
isn't in the Image itself), so we can use that to be sure we
place the initrd sufficiently far into the RAM.

Patch 1 in this series adjusts our "where do we put the initrd"
heuristic so that it always places it at least after whatever
our best guess at the kernel size is. (This might still not
be right for images like self-decompressing 32-bit kernels, where
there's no way to know how big the kernel will be after
decompression.) Patch 2 makes load_aarch64_image() return the
kernel size as indicated in the Image file header, so that for
the specific case of AArch64 Image files we will definitely not
put the initrd on top of them.

I've given this a quick smoke test but I don't have a very large
Image kernel to hand, so testing appreciated.

thanks
-- PMM

Peter Maydell (2):
  hw/arm/boot: Avoid placing the initrd on top of the kernel
  hw/arm/boot: Honour image size field in AArch64 Image format kernels

 hw/arm/boot.c | 51 +++++++++++++++++++++++++++++++++++----------------
 1 file changed, 35 insertions(+), 16 deletions(-)

-- 
2.20.1

On Fri, May 03, 2019 at 06:13:45PM +0100, Peter Maydell wrote:
> This patchset attempts to fix https://bugs.launchpad.net/qemu/+bug/1823998
> which reports that we don't handle kernels larger than 128MB
> correctly, because we allow the initrd to be placed over the
> tail end of the kernel. AArch64 kernel Image files (since v3.17)
> report the total size they require (including any BSS area that
> isn't in the Image itself), so we can use that to be sure we
> place the initrd sufficiently far into the RAM.
> 
> Patch 1 in this series adjusts our "where do we put the initrd"
> heuristic so that it always places it at least after whatever
> our best guess at the kernel size is. (This might still not
> be right for images like self-decompressing 32-bit kernels, where
> there's no way to know how big the kernel will be after
> decompression.) Patch 2 makes load_aarch64_image() return the
> kernel size as indicated in the Image file header, so that for
> the specific case of AArch64 Image files we will definitely not
> put the initrd on top of them.
> 
> I've given this a quick smoke test but I don't have a very large
> Image kernel to hand, so testing appreciated.

I've just given this a go with three very large images built from v5.1:

* ~113.6MiB raw, ~134.5MiB effective
* ~131.0MiB raw, ~152.0MiB effective
* ~225.6MiB raw, ~247.0MiB effective

Prior to these patches (with pristine QEMU commit c56247e55bde4386) both
would silently fail to boot, and with these patches applied both all
three begin booting and produce console output. The first two get to
userspace, and the third crashes due to an unrelated Linux bug.

So FWIW:

Tested-by: Mark Rutland <mark.rutland@arm.com>

Thanks for putting this together!

Thanks,
Mark.